Image Segmentation
Johan Lie1, Marius Lysaker2, and Xue-Cheng Tai1 1 Department of Mathematics, University of Bergen,
Joh. Brunsgate 12, N-5008 Bergen, Norway
{johanl, tai}@mi.uib.no http://www.mi.uib.no/BBG http://www.mi.uib.no/˜tai
2 Simula Research Laboratory, M. Linges v 17, Fornebu P.O.Box 134, N-1325 Lysaker, Norway
Abstract. In this work we discuss variants of a PDE based level set method. Traditionally interfaces are represented by the zero level set of continuous level set functions. We instead use piecewise constant level set functions, and let interfaces be represented by discontinuities. Some of the properties of the standard level set function are preserved in the proposed method. Using the methods for interface problems, we mini-mize a smooth locally convex functional under a constraint. We show numerical results using the methods for image segmentation.
1
Introduction
The level set method was proposed by Osher and Sethian in [19] as a versatile tool for tracing interfaces separating a domainΩinto subdomains. Interfaces are treated as the zero level set of higher dimensional functions. Moving the inter-faces can implicitly be done by evolving level set functions instead of explicitly moving the interfaces. Applications of the level set method include image anal-ysis, reservoir simulation, inverse problems, computer vision and optimal shape design [5, 4, 10, 26, 18, 22, 24]. In this work, we discuss some variants of the level set method. The primary concern for our approach is to remove the connection between the level set functions and the signed distance function and thus re-move some of the computational difficulties associated with the calculation of the Eikonal equation. Another motivation is to avoid numerical problems associ-ated with the Heaviside and Delta functions used in some level set formulations [5, 25]. The third concern of this approach is to develop fast algorithms for level set methods. Due to the fact that the functional and the constraints for this approach are rather smooth, it is possible to apply Newton types of iterations to construct fast algorithms for the proposed model. For experimental purposes, we have used gradient type of methods here, and we restrict ourselves to image segmentation.
R. Kimmel, N. Sochen, J. Weickert (Eds.): Scale-Space 2005, LNCS 3459, pp. 573–584, 2005. c
For a given digital imageu0:Ω→R, the aim is to separateΩ into a set of subdomainsΩi such that Ω=∪ni=1Ωi andu0 is nearly a constant in each Ωi. Having determined the partition ofΩ into a set of subdomainsΩi, one can do further modelling on each domain independently and automatically. One general image segmentation model was proposed by Mumford and Shah in [16]. For numerical approximations, see [2]. Using this model, the imageu0is decomposed intoΩ=∪iΩi∪Γ, where Γ is a curve separating the different domains. Inside each Ωi, u0 is approximated by a smooth function. The optimal partition of Ω is found by minimizing the Mumford-Shah functional. Chan and Vese [5, 25] minimized the piecewise constant Mumford-Shah functional using level set functions. Motivated by the Chan-Vese approach, we will in this article solve the segmentation problem in a different way, i.e. by introducing a piecewise constant level set functionφ. Instead of using the zero level of a function to represent the interface between subdomains, the interfaces are implicitly represented by the discontinuities of a set of characteristic functions ψi. Note that both the Chan-Vese model and our model can be extended to shape recognition using the framework of [7, 6].
The rest of this article is structured as follows. Our framework and the min-imization functional used for image segmentation is formulated in§2. The seg-mentation problem is formulated as a minimization problem with a smooth cost functional under a constraint. We are minimizing the piecewise constant Mumford-Shah functional associated with special level set models. In§3 and§4 we explain our two variants of the level set method for image segmentation in more detail. Both sections include algorithms and numerical results. We con-clude with a brief discussion. For a more detailed treatment of the methods, including more numerical results we refer the reader to [14, 15]. We also refer the reader to [23, 21, 9, 11, 7] for some related methods.
2
A Framework for Subdomain Representation
In this section a framework for representing subdomains ofΩis developed. Each subdomainΩiis associated with a basis functionψi, such thatψi = 1 inΩiand zero elsewhere. The basis functions are constructed using one or several level set functions{φj}l
j=1. Two different realizations of the basis functionsψi are shown in §3 and §4. Each ψi is compactly supported in Ωi. Thus we can construct a piecewise constant functionuby a weighted sum of the characteristic functions. If we letc={ci}ni=1be a set of real scalars, we can represent a piecewise constant functionutaking thesendistinct constant values by
u= n i=1
ciψi. (1)
Letη be Gaussian noise, andu0 = u+η. If u0 is almost equal to a constant in n subdomains it can be approximated by (1) provided we optimally choose c and {ψi}n
constraint corresponding to the choice of basis functions. The constraint con-trols the structure of possible solutions. We return to the specific constraints in §3 and§4.
The simple structure of the characteristic functions gives us the opportunity to measure the lengths of curves surroundingΩiand the area of each regionΩiby
|∂Ωi|= Ω
|∇ψi|dx, and |Ωi|= Ω
ψidx. (2)
Here we note that|∂Ωi|is the Total Variation (TV)-norm ofψi [20].
The above framework can be used as a tool for image segmentation. Letu0be an image to be segmented. We want to construct a piecewise constant function uwhich approximatesu0in a proper sense. The segmentation can be formulated as a minimization of the following functional
F(φ,c) =1 2 Ω |u−u0|2dx+β n i=1 Ω |∇ψi|dx, (3)
whereuis on the form (1). The first term of F is a least squares fidelity term, measuring the closeness ofutou0. The second term is a regularizer term mea-suring the length of all the curves separating the subdomains. We here note the similarity between functional (3) and the functional used by Chan and Vese in [25], whereu is represented using Heaviside functions. For uniquely classifying each point in the image, we need to introduce a constraintK(φ) = 0 (c.f.§3§4) and solve the constrained optimization problem
min
c,φ F(c, φ) subject to K(φ) = 0. (4) This problem is solved using the augmented Lagrangian method [1, 17]. A mini-mizer ofFcorresponds to a saddle-point of the augmented Lagrangian functional
L(c, φ, λ) =F(c, φ) + Ω λK(φ)dx+r 2 Ω |K(φ)|2dx, (5)
whereλis a function defined on the same domain asφcalled the Lagrangian mul-tiplier, r ∈R+ is a penalty parameter being characteristic for the augmented Lagrangian method. r can either be chosen to be a fixed small number dur-ing the iterative process or it can be increased durdur-ing iterations to accellerate convergence [17]. We have used both the augmented Lagrangian method and the standard Lagrangian method. Numerical experiments indicate that the aug-mented Lagrangian method is best suited for this minimization problem. At a saddle-point of (5) we must have
∂L ∂φ = 0, ∂L ∂ci = 0 and ∂L ∂λ = 0. (6)
Essentially we minimizeLw.r.tcandφ, and maximizeLw.r.tλ. In§3 and§4 we introduce iterative algorithms to find the saddle-points in (6) coming from
two different level set formulations. The rest of the current section is devoted to calculations being common for both approaches.
Sinceuis linear with respect toc, we see thatLis quadratic with respect to c. Thus the minimization problem w.r.tccan be solved exactly. Note that
∂L ∂ci = Ω ∂L ∂u ∂u ∂ci = Ω (u−u0)ψidx, for i= 1,2, . . . n. (7)
Therefore, the minimizer satisfies a linear system of equations Ack =b in the following form n j=1 Ω (ψjψi)cki dx= Ω u0ψidx, for i= 1,2, . . . n. (8)
In the aboveψj =ψj(φk),ψi=ψi(φk) and thus,ck ={cki}ni=1 depends onφk. We form the (n×n) matrix A and vector b and solve the equationAck = b using an exact solver. The minimization with respect toφwill be solved by the following gradient method
φnew=φold−∆t∂L ∂φ(c, φ
old, λ), (9) where∆tis a small positive number determined by trial and error. For a givenc andλ, we need to iterate many times in order to find the minimizer with respect toφ. This can be interpreted as a scalespace method withtas the scale.
3
Piecewise Constant Level Set Method Using a
Polynomial Approach
We first present the polynomial piecewise constant level set method (PCLSM). Assume that we need to find n regions {Ωi}ni=1 which form a portion of Ω. In order to identify the regions, we want to find a piecewise constant function taking the values
φ=iin Ωi, i= 1,2, . . . , n. (10) With this approach we just need one function to identify all the phases in Ω. The basis functionsψi associated withφare defined in the following form
ψi = 1 αi n j=1 j=i (φ−j) and αi= n k=1 k=i (i−k). (11)
It is clear that the functionugiven by (1) is a piecewise constant function and u= ci in Ωi ifφ is as given in (10). The function u is a polynomial of order n−1 inφ. Eachψi is expressed as a product of linear factors of the form (φ−j), with the ith factor omitted. Thereupon ψi(x)=1 for x ∈ Ωi, and ψi(x = 0
elsewhere as long as (10) holds. To ensure that equation (1) gives us a unique representation ofu, i.e. at convergence different values of φ should correspond to different function valuesu(φ) in (1), we introduce
K(φ) = (φ−1)(φ−2)· · ·(φ−n) = n i=1
(φ−i). (12) If a given functionφ:Ω→R satisfies
K(φ) = 0, (13)
there exists a unique i ∈ {1,2, . . . , n} for every x ∈ Ω such that φ(x) = i. Thus, each point x ∈ Ω can belong to one and only one phase if K(φ) = 0. The constraint (13) is used to guarantee that there is no vacuum and overlap between the different phases. In [27] some other constraints for the classical level set methods were used to avoid vacuum and overlap.
Following the framework in§2, we will use the basis functions (11), the con-straint (12) and the representation (1) ofu. To find a minimizer of (5), we need to find the saddle point where ∂L∂φ = 0 and ∂L∂λ = 0. Remember that ∂c∂L
i is zero if{ci}n
i=1are computed from (8). We use the Uzawa-type Algorithm 1 to find a saddle point of L(c, φ, λ). The algorithm has a linear convergence rate and its convergence has been analyzed in [13] and used in [4, 3].
Algorithm 1. Choose initial values forφ0 and
λ0. For k= 1,2, . . ., do: – Findck from L(ck, φk−1, λk−1) = min c L(c, φ k−1 , λk−1). (14)
– Use (1) to updateu=ni=1ckiψi(φk−1).
– Findφk from
L(ck, φk, λk−1) = min φ L(c
k, φ, λk−1)
. (15)
– Use (1) to updateu=ni=1ckiψi(φk).
– Update the Lagrange-multiplier by
λk =λk−1+rK(φk). (16)
– If not converged: Set k=k+1 and go to step 1. To compute dLdφ we utilize the chain rule to get
∂L ∂φ = (u−u0) ∂u ∂φ−β n i=1 ∇·|∇∇ψi ψi| ∂ψi ∂φ +λ ∂K ∂φ +rK ∂K ∂φ. (17) It is easy to get ∂u/∂φ, ∂ψi/∂φ and ∂K/∂φ from (1), (11) and (12). We use the gradient method (9) to solve (15). We do a fixed number of iterations, for example 400 iterations or stop the iteration after theL2 norm of gradient has been reduced by 10%.
Remark 1. The updating for the constant values in (14) is ill-posed. A small perturbation of the φ function produces a large perturbation for the ci val-ues. Due to this reason, we have tried out a variant of Algorithm 1. In each iteration we alternate between (15) and (16), while (14) is only carried out if K(φnew)
L2< 101K(φold)L2. Here,φolddenotes the value ofφwhen (14) was carried out the last time andφnewdenotes the current value ofφ. If we use such a strategy, we can do just one or a few iterations for the gradient scheme (9) and Algorithm 1 is still convergent. This strategy is particular efficient when the amount of noise is high.
Remark 2. In Algorithm 1, we give initial values forφandλ. We first minimize with the constant values, and then minimize with the level set function. The multiplier is updated in the end of each iteration. In situations where good initial values forc are available, an alternative variant of Algorithm 1 may be used, i.e. we first minimize with the level set function followed by a minimization for the constant values and then update the multiplier.
Remark 3. There is no particular reason why we chose to use integers as roots of the polynomial. Perhaps a better choice would be to use a Chebyshev polynomial with its roots as buildingblocks for the characteristic functionsψi. This could for example be accomplished by interchanging the integer values in (10) by the roots of a Chebyshev polynomial of degreendefined on the interval [−1,1] [12].
zi= cos π(i−12) n in Ωi, i= 1,2, . . . , n. (18) The exact same framework could then be used to construct a set of characteristic functions{ψi}n i=1 by ψi= 1 αi n j=1 j=i (φ−zj) and αi= n k=1 k=i (i−zk). (19)
The corresponding constraint would then be
K(φ) = (φ−z1)(φ−z2)· · ·(φ−zn) = n i=1
(φ−zi). (20)
In this work we have not implemented this approach.
3.1 Numerical Experiments Using the Polynomial Approach
In this section we validate the piecewise constant level set method with numerical examples. We only consider 2-D images and restrict ourself to gray-scale images, although the model permits any dimension and can be extended to vector-valued images as well. Our results will be compared with the related works [5, 25]. To complicate the segmentation process we typically expose the original image with
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100 0 20 40 60 80 100 0 20 40 60 80 100 0 1 2 3 4 5 (a) (b) Fase 1 20 40 60 80 100 20 40 60 80 100 Fase 2 20 40 60 80 100 20 40 60 80 100 Fase 3 20 40 60 80 100 20 40 60 80 100 Fase 4 20 40 60 80 100 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100 0.5 1 1.5 2 2.5 3 3.5 4 4.5 (c) (d)
Fig. 1.(a) Observed imageu0 (SNR≈5.2). (b) Initial level set function φ. (c) Each separate phaseφ= 1∨2∨3∨4 are depicted as a bright region. (d) At convergenceφ approaches 4 constant values
Gaussian distributed noise and use the polluted image as the observation datau0. To demonstrate a 4-phase segmentation we begin with a noisy synthetic image containing 3 objects (and the background) as shown in Figure 1(a). This is the same image as Chan and Vese used to examine their multiphase algorithm [5, 25]. The observation data u0 is given in Figure 1(a) and the only assumption we make is that a 4-phase model should be utilized to find the segmentation. In Figure 1(d) theφfunction is depicted at convergence. The functionφapproaches the predetermined constantsφ= 1∨2∨3∨4. Each of these constants represents one unique phase as seen in Figure 1(c). Our result is in accordance with what Chan and Vese reported in [5, 25].
In many applications the number of objects to detect are not known a pri-ori. A robust and reliable algorithm should find the correct segmentation even when the exact number of phases is not known. By introducing a model with
Fase 1 20406080100 20 40 60 80 100 Fase 2 20406080100 20 40 60 80 100 Fase 3 20406080100 20 40 60 80 100 Fase 4 20406080100 20 40 60 80 100 Fase 5 20406080100 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 1000 1 2 3 4 5 6 (a) (b)
Fig. 2. (a) Each phaseφ = 1∨2∨3∨4∨5 is depicted as a bright region. (b) At convergenceφapproaches 4 constant values
(a) (b) (c)
(d) (e)
Fig. 3.Character and number segmentation from a car plate using our method (b),(c) and CV (d),(e)
more phases than one actually needs, we can find the correct segmentation if all superfluous phases are empty when the algorithm has converged. To see if our algorithm can handle such a case we again use Figure 1(a) as the observation image and utilize a 5-phase model. Our results are reported in Figure 2. One of the 5 phases must be empty if a 5-phase model is used to find a 4-phase segmen-tation. Due to the high noise level some pixels can easily be misclassified and contribute to the phase that should be empty. (The regularization parameterβ is not tuned in order to get rid of this misclassification.) The level set function shown in Figure 2(b) approaches the constantsφ= 1∨2∨4∨5, except from the few misclassified pixels whereφ = 3 as seen in Figure 2(a). By comparing Figure 1(c) (where a 4-phase model is used) and Figure 2(a) (where a 5-phase model is used), we observe only small changes in the segmented phases, except from the extra nonempty phaseφ= 3 in Figure 2(a).
Below we proceed with one example using a real image. We want to demon-strate that PPCLSM (polynomial PCLSM) can be uses to extract characters or numbers from images. We use an image of a license plate. To evaluate the seg-mentation process, the Chan/Vese method [5, 25], for short (CVM), is examined using the same input image. Both algorithms are processed using two different regularization parameters. With the amount of noise in Fig. 3 (a), both PPCLSM and CVM miss some details along the edges of the characters and numbers. To compare the result of (CV) and (PPCLSM), both algorithms were terminated after 500 iterations. The regularization parameters used in the experiment are (b)β = 0.1, (c) β = 3, (d)β = 1 and (e) β = 104. Thus we observe that both our method and CVM are quite robust to noise, and the choice of regulariza-tion parameters. The processing time is almost equal for both methods in this example. We have done no efforts for code optimization. The complexity of the implementation is also almost equal for both methods.
4
The Binary Approach for PCLSM
We will now introduce an alternative realization of the characteristic functions in (1). Using the following approach, we can represent a maximum of 2N sub-domains, usingN level set functions{φi}Ni=1. To simplify notation, we form the vectorφ={φ1, φ2, . . . , φN}. Let us first assume that the interfaceΓ is enclosing Ω1⊂Ω. By standard level set methods the interior ofΩ1is represented by points x : φ(x)>0, and the exterior ofΩ1 is represented by pointsx : φ(x)<0. We instead let φ(x) = 1 if x is an interior point of Ω1 and φ(x) = −1 if xis an exterior point ofΩ1. As proposed,Γ is implicitly defined as the discontinuity of φ. Representing four subdomains is done analogously as in [25, 7] by
u(x) = ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ c1,if φ1(x) = 1, φ2(x) = 1, c2,if φ1(x) = 1, φ2(x) =−1, c3,if φ1(x) =−1, φ2(x) = 1, c4,if φ1(x) =−1, φ2(x) =−1.
Thus, a piecewise constant function taking four different constant values can be written u= c1 4(φ1+ 1)(φ2+ 1)− c2 4(φ1+ 1)(φ2−1) −c3 4(φ1−1)(φ2+ 1) + c4 4(φ1−1)(φ2−1). (21) Using (21), we can form the set of basis functionsψi as in the following
u=c1 1 4(φ1+ 1)(φ2+ 1) ψ1 +c2(−1) 1 4(φ1+ 1)(φ2−1) ψ2 +. . . , (22)
and we can write:u=4i=1ciψi. The set ofψifor the multiple subdomain case is constructed by generalization of this. Fori= 1,2, . . . ,2N, let (bi−1
1 , bi− 1
2 , . . . , bi− 1 N ) be the binary representation ofi−1, wherebi−j 1= 0∨1.Lets(i) =Nj=1bi−j 1, and writeψi anduas
ψi=(−1)2Ns(i)Nj=1(φj+ 1−2bi−j 1) and u= 2N i=1
ciψi. (23) It is now easy to see that these basis functions have the properties needed for the framework in§2. Using this representation for the basis functions, we need N constraints, one constraint Ki to each of the level set functions φi. We use the constraints:Ki(φi) =φ2i−1∀i. SettingKi(φi) = 0 impliesφi can only take the values±1 at convergence.
Having determined the choice of basis functions{ψi}ni=1 and the representa-tion of uby (23), we find the saddle point of Lby the augmented Lagrangian method. This means that we must minimizeLw.r.tφ andc, and maximizeL w.r.tλ, which has the same dimension asφ.
We minimizeLw.r.tφby using the gradient method (9) for all theN level set functions. The gradients for the level set functions are given as
∂L ∂φi= (u−u0) 2N j=1 cj∂ψj ∂φi −β 2N j=1 ∇·|∇∇ψj ψj| ∂ψj ∂φi +2λiφi+ 2r(φ2i −1)φi. (24) The constraintsKi are independent of the constant valuesciand thus the same formula (8) can be used to update theci values.
Similar to the algorithm used for the polynomial approach for the PCLSM, we use the following algorithm to find a saddle point for the binary approach for the PCLSM.
Algorithm 2. Choose initial values forφ0andλ0. Fork= 1,2, . . ., do:
– Updateφk by (9), to approximately solve L(ck−1,φk,λk−1) = min φ L(c k−1 ,φ,λk−1). (25) – Constructu(ck−1,φk) by u=2i=1N ck−i 1ψki. – Updateck by (8), to solve L(ck,φk,λk−1) = min c L(c,φk,λk−1). (26)
– Update the multiplier by
λk=λk−1
+rK(φk). (27)
– If not converged: Set k=k+1 and go to step 1.
Remark 4. Remarks 1 and 2 after Algorithm 1 also apply to algorithm 2.
(a) (b) (c)
Fig. 4. In this example, the image (a) is first segmented using the isodata method (b). Then the result is further processed using our method, and the final result is shown in (c)
4.1 Numerical Experiments Using the Binary Approach
Now we will present one of the numerical results achieved using the binary piece-wise constant level set formulation (BPCLSM). For more numerical results, see [15]. We will here show a segmentation of a Magnetic Resonance image of a brain using two level set functions. The goal is to partition the image into three differ-ent tissue classes in addition to the background. To accelerate the convergence of our method, we first preprocess the image using a simple tresholding, the iso-data method [8]. In (b), we show the result of the isoiso-data segmentation ofu0[8]. We here observe that main structures are preserved, but also highly oscillating patterns occur. We use the results from the isodata segmentation to construct initial values forφ1andφ2, run our algorithm with these initial values, and end up with the image depicted in Fig. 4 (c). Observe that the main structures are still very well preserved, but most of the (unwanted) highly oscillating patterns are removed. By initializing the algorithm in this way, we both accelerate the convergence of our algorithm, and in addition more or less avoid the problem of local minimizers. In the example, we used the parameters β = 5·10−3,
µ= 5, and 1000 iterations.
5
Conclusion
In this article we have discussed a framework for subdomain identification. We have also pointed out two methods for image segmentation using this framework. Recently, there has been done work to extend the method to 3-D data sets. Work is also done to incorporate a Newton-type of iteration for improving the convergence properties of the method. The PPCLSM is favourable in terms of computational complexity and memory requirements, and in terms of handling cases where a priori information of the number of subdomains is lacking. Molding BPCLSM into existing software for level set methods is possibly easier than PPCLSM because of similarities in the machinery of standard level set methods.
References
1. D. P. Bertsekas, Constrained optimization and Lagrange multiplier
meth-ods,Academic Press Inc.,1982.
2. A. Chambolle,Image segmentation by variational methods: Mumford and Shah
functional and the discrete approximations, SIAM J. Appl. Math., 55 (1995), pp. 827–863.
3. T. F. Chan and X.-C. Tai,Identification of discontinuous coefficients in elliptic
problems using total variation regularization, SIAM J. Sci. Comput., 25 (2003), pp. 881–904.
4. T. F. Chan and X.-C. Tai,Level set and total variation regularization for elliptic
inverse problems with discontinuous coefficients, J. Comput. Phys., 193 (2003), pp. 40–66.
5. T. F. Chan and L. A. Vese, Active contours without edges, IEEE Trans. Im.
6. D. Cremers, T. Kohlberger, and C. Schn¨orr,Shape statistics in kernel space
for variational image segmentation, Patt. Recogn., 36 (2003), pp. 1929–1943. 7. D. Cremers, N. Sochen, , and C. Schn¨orr, Multiphase dynamic labeling for
variational recognition-driven image segmentation, in ECCV 2004, LNCS 3024, 2004, pp. 74–86.
8. F. R. Dias Velasco,Thresholding using the ISODATA clustering algorithm, IEEE
Trans. Systems Man Cybernet., 10 (1980), pp. 771–774.
9. S. Esedoglu and Y.-H. R. Tsai,Threshold dynamics for the piecewise constant
mumford-shah functional, UCLA-CAM Report 04-63, (2004).
10. R. P. Fedkiw, G. Sapiro, and C.-W. Shu, Shock capturing, level sets, and
PDE based methods in computer vision and image processing: a review of Osher’s contributions, J. Comput. Phys., 185 (2003), pp. 309–341.
11. F. Gibou and R. Fedkiw,A fast hybrid k-means level set algorithm for
segmen-tation, Stanford Tech. Rep., 2002, (in review)., (2002).
12. M. T. Heath,Scientific computing: an introductory survey., 2001.
13. K. Kunisch and X.-C. Tai,Sequential and parallel splitting methods for bilinear
control problems in Hilbert spaces, SIAM J. Numer. Anal., 34 (1997), pp. 91–118. 14. J. Lie, M. Lysaker, and X.-C. Tai,A variant of the level set method and
appli-cations to image segmentation, UCLA, CAM-report, 03-50, (2003).
15. J. Lie, M. Lysaker, and X.-C. Tai,A binary level set model and some
applica-tions for mumford-shah image segmentation, UCLA, CAM-report, 04-31, (2004). 16. D. Mumford and J. Shah,Optimal approximation by piecewise smooth functions
and associated variational problems, Comm. Pure Appl. Math, 42 (1989), p. 577685. 17. J. Nocedal and S. J. Wright,Numerical optimization, Springer Series in
Op-erations Research, 1999.
18. S. Osher and R. Fedkiw, Level Set Methods and Dynamic Implicit Surfaces,
vol. 153 of Appl. Math. Sci., Springer, 2003.
19. S. Osher and J. A. Sethian,Fronts propagating with curvature-dependent speed:
algorithms based on Hamilton-Jacobi formulations, J. Comput. Phys., 79 (1988), pp. 12–49.
20. L. Rudin, S. Osher, and E. Fatemi,Nonlinear total variation based noise
re-moval algorithm, Physica D., 60 (1992), pp. 259–268.
21. C. Samson, L. Blanc-Feraud, G. Aubert, and J. Zerubia,A level set model
for image classification, IJCV, 40 (2000), pp. 187–198.
22. J. A. Sethian,Level set methods and fast marching methods,Cambridge University
Press, second ed., 1999.
23. B. Song and T. F. Chan,A fast algorithm for level set based optimization, Tech.
Rep. CAM 02-68, UCLA, 2002.
24. X.-C. Tai and T. F. Chan,A survey on multiple set methods with applications
for identifying piecewise constant functions, Int. J. Num. Anal. and Mod., 1 (2004), pp. 25–48.
25. L. A. Vese and T. F. Chan, A multiphase level set framework for image
seg-mentation using the mumford and shah model, Int. J. of Comp. Vis., 50 (2002), pp. 271–293.
26. J. Weickert and G. K¨uhne,Fast methods for implicit active contour models,
in Geometric level set methods in imaging, vision, and graphics, Springer, 2003, pp. 43–57.
27. H.-K. Zhao, T. Chan, B. Merriman, and S. Osher, A variational level set
